Disclaimer: This article appeared in the AIMA Journal (Dec 2004), which is published by The Alternative Investment

Management Association Limited (AIMA). No quotation or reproduction is permitted without the express written
permission of The Alternative Investment Management Association Limited (AIMA) and the author. The content of this
article does not necessarily reflect the opinions of the AIMA Membership and AIMA does not accept responsibility for
any statements herein.

High frequency automated trading

By Andy Webb, Omega Derivatives

The concept of automated trading has attracted rapidly growing interest in recent years. In certain
markets, such as exchange-traded futures, it has already become an everyday fact of life. In others, such
as interbank spot FX, the party is just beginning to get underway.

Irrespective of the specific market, a focus for this growing interest has been high frequency
autotrading. While technological advances have certainly been a factor in this, another major driver has
been the inexorable decline in transaction costs. This reduces the minimum acceptable average profit
per trade, thus allowing trade frequency to be increased without degrading overall profitability.

Spreading the risk, smoothing the curve

One of the most important advantages of high frequency autotrading is ease of diversification. While
diversification by market is widespread, diversification by timeframe and trading model have historically
been less common. A high frequency automated trading environment makes it relatively straightforward
to deploy the same (or differing) trading models across a portfolio of timeframes.

The benefits of time diversification in terms of reducing correlation (even when using the same basic
trade system rules) can be striking. Figure 1 shows the equity curve for a very simple reversal trading
system applied to 20-minute GBPUSD data over twenty trading sessions. (Twelve hour trading sessions,
0600 to 1800 UK time). Applying the same trading logic in a 1-minute timeframe obviously generates a
far higher trade count than the 20-minute data (approx 5000 trades versus 270). If the 1-minute trades
closest in time to the 20-minute trades are sampled to produce two sets of results of the same size, the
correlation between the two data sets is just over 0.1 – a small positive correlation.

Although the trade system in Fig 1 is profitable, it has two notable drawdowns at “A” and “B”, and an
underwater period at “C” immediately after trading commences. However, by virtue of the larger
number of data points it generates, high frequency trading can be added to the mix to improve the
smoothness of the longer time frame’s equity curve and to reduce its standard deviation of return.
Figure 2 shows the effect of adding the 5000 results for the 1-minute time frame mentioned above to the
20-minute time frame results in Figure 1. In addition to largely removing the drawdowns at letters A and
B, the higher frequency results also eradicate the initial underwater period at letter C. The standard
deviation of return was also reduced by a factor of seven.

Fig 1
Disclaimer: This article appeared in the AIMA Journal (Dec 2004), which is published by The Alternative Investment
Management Association Limited (AIMA). No quotation or reproduction is permitted without the express written
permission of The Alternative Investment Management Association Limited (AIMA) and the author. The content of this
article does not necessarily reflect the opinions of the AIMA Membership and AIMA does not accept responsibility for
any statements herein.

Fig 2

Virgin territory…
A further advantage of autotrading when applied to very short time frames at high frequency is that it is
operating in what is (relatively speaking) virgin territory. The physical limitations of human traders and
the fact that fully automated FX trading is not universal, means that the shortest timeframes (such as
individual price ticks) still contain a certain amount of non-random behaviour. At this high-speed
microstructure level, the market is not yet saturated with multiple participants all chasing the same
trading opportunities.

Figure 3 gives an indication of the sort of non-random bias that can exist at the tick-by-tick level. The
“Up” column shows the percentage frequency of the market making an up tick after two consecutive
down ticks and the “Down” column the percentage frequency of a down tick after two consecutive up
ticks. The results are derived from testing in 4000 tick blocks at various dates, with each pattern
occurring between approximately 400 and 600 times during each sample period. “Flat” ticks
(consecutive ticks at the same price) were ignored for the purposes of pattern recognition.

Up Down
EURUSD 63.8 63.4
USDJPY 57.9 58.5
EURCHF 59.1 58.3

Fig 3

At first glance this degree of bias seems to offer an exceptional opportunity for autotrading. However,
closer inspection reveals an important caveat that is generally applicable to high frequency trading: will
an order based on these patterns actually be executable in real time? For example, a bid placed at the
same price as the last down tick will be at the back of the order queue. Unless there are several
subsequent flat ticks at the bid price, the order is highly unlikely to be filled. Therefore when testing
such a trading model it is essential to have access to historical volume information at a tick-by-tick
level, which can be used to gauge the probability of a fill when back testing.

In addition to non-random opportunities, high frequency trading also allows existing data to be used in a
slightly different way. One avenue of analysis that is often under-explored in slower frequency trading is
depth of market data. For example, analysing the percentage of trades in the last fifteen seconds that
have been conducted at the bid and ask and comparing that with current market depth can offer a
Disclaimer: This article appeared in the AIMA Journal (Dec 2004), which is published by The Alternative Investment
Management Association Limited (AIMA). No quotation or reproduction is permitted without the express written
permission of The Alternative Investment Management Association Limited (AIMA) and the author. The content of this
article does not necessarily reflect the opinions of the AIMA Membership and AIMA does not accept responsibility for
any statements herein.

useful indication of short-term market direction. However, this type of method is only applicable to
depth of market data close to the current trading level in highly active markets. In other circumstances
the data tends to be distorted by spoof orders that are “good until close”.

Increasing volume…
High frequency autotrading obviously also benefits the market as a whole by improving liquidity. The
record volumes posted by many futures exchanges last year have been widely attributed to its wider
adoption. A similar situation applies elsewhere – for example EBS has also seen evidence of the volume
potential during the initial trial of its ATI (Automated Trading Interface). All the banks participating in
the trial appreciably increased their daily volume on EBS, with the increase in some cases being very
substantial.

Much existing autotrading activity is not so much due to traders’ attempts to capture directional moves
in the market, but to various forms of arbitrage activity. For example, one London based non-bank
trading operation that automatically arbitrages inefficiencies in exchange matching engines frequently
accounts for 20-30,000 contracts per day in CME currency futures. During one period last year it was also
responsible for more than 15% of daily volume in the CBOT Treasury Bond future.

…but increasing traffic as well

Gratifying though this surging volume is for both exchanges and participants, there is a downside: order
traffic. Minimising order volume is not usually a priority for those building an automated trading system,
which can result in models that fire literally hundreds of orders per second at the market. An
appreciable percentage of these messages may not even be strictly necessary for the profitable
operation of the model. This surfeit of orders can cause major problems for exchanges in terms of
bandwidth and matching engine capacity.

An additional factor is that the proportional growth in order traffic due to autotrading systems is vastly
exceeding the growth in actual trading volume. For example, Euronext.liffe’s average daily futures
volume on LIFFE has grown roughly 50% for all contracts since 2001 – but the order messages sent to the
matching engine have risen by 300% in the same period.

Exchanges have responded to this by enhancing their infrastructure, but also by applying controls. Some
have even instituted fines for participants who generate order messages above a specified multiple of
their filled orders. More sophisticated development environments can largely negate this problem
through aggregation and netting of orders. (For example substituting a single ten lot order for ten
individual one lot orders.)

The need for speed - and synchronicity

In a machine versus machine trading environment outright speed is obviously crucial, but there are other
equally critical but less obvious issues. One example is the manner in which events are handled within
the same trading system or within multiple systems operating as a portfolio. Historically, especially
when dealing with slower CPUs, there has been bias towards analytical software operating on an
asynchronous basis. Under this model, the output of individual program threads is not kept synchronised,
so for example one indicator in a chart window might update before another. This has the advantage of
allowing efficient processor loading, and is usually of little concern to a human trader.

This is certainly not the case in an autotrading environment. Consider multiple trading systems applied
to multiple markets, where the risk management logic requires that before a new position is opened a
check is made to ensure that there is not an excessive correlation with positions already open. For this
check to be accurate, all systems must be synchronised before it is carried out. Therefore, when a global
event1 occurs (such as incoming data tick) all code relating to that tick in all the systems must have

1
In this instance, an event that affects all the systems.
Disclaimer: This article appeared in the AIMA Journal (Dec 2004), which is published by The Alternative Investment
Management Association Limited (AIMA). No quotation or reproduction is permitted without the express written
permission of The Alternative Investment Management Association Limited (AIMA) and the author. The content of this
article does not necessarily reflect the opinions of the AIMA Membership and AIMA does not accept responsibility for
any statements herein.

finished executing before the correlation check can be made. A development environment that supports
this synchronous method of operation is therefore essential if an autotrading system is to function in real
time as originally intended. Otherwise the developer has to write additional code that checks that all
components of a system are equally up to date before summarising and placing (or not placing) an order.
Apart from the extra work, this increases the risk of errors.

A complement, not a replacement

The opportunities offered by autotrading have prompted some to predict that it will rapidly make the
human trader redundant. However, this looks unlikely to happen in the very immediate future. While
significant progress has been made in simulating human intelligence, automated trading engines still fall
some way short of human traders when it comes to finessing large or awkward orders. Those already
using automated trading have discovered that using it to replicate the human skill of finessing an order is
not easy. As a result, their preference is to use autotrading to complement rather than replace human
traders by using it for low value tactical trading or for high frequency activity.